Most of my work starts with someone handing me a folder of files and saying 'we don't really know what's in here.' Audio recordings with no transcripts. Research papers that contradict each other. Sensor logs from a lab experiment that ran for six months.

I figure out what's actually there, build something that can learn from it, and get it into production. Not just a notebook that demos well. That last part matters more than people admit.

Lately I've been working a lot with audio and unstructured text. I've trained transformer models on datasets I had to build myself because nothing off-the-shelf existed. I've also written simulation code that needed to match real lab results, which is a very different pressure than matching a benchmark.

I'm not the person who'll sell you on a technique before understanding your problem. I'd rather spend an extra week getting the framing right than move fast and build the wrong thing.